Author

Mark McDonnell, mark.mcdonnell@unisa.edu.au

Introduction

This repository contains code aligned with an entry to DCASE 2020 Challenge Task 1B.

Information on the DCASE challenge can be found here: http://dcase.community/challenge2020/task-acoustic-scene-classification

A technical report on the challenge entry can be found at https://github.com/McDonnell-Lab/DCASE2020_Task1b/blob/master/TechnicalReport/McDonnell_USA_Task1b.technical_report.pdf

Installation

For this repository (https://github.com/McDonnell-Lab/DCASE2020_Task1b), the following are strongly recommended:

1. Ubuntu >= 16.04
2. a working high end GPU with >= 11 GB RAM and drivers capable of running tensorflow 1.13 or higher
3. use of Anaconda
4. use of jupyter notebooks

Example instructions for installing anaconda are here: https://www.digitalocean.com/community/tutorials/how-to-install-the-anaconda-python-distribution-on-ubuntu-16-04

This repository contains a script that can automatically download and unzip the Task 1b challenge data.

You must have anaconda installed on linux for the script to work.

For example, first download this repositor, e.g. go to https://github.com/McDonnell-Lab/DCASE2020_Task1b and click on "Download zip" then copy to your desired location and unzi.

Or on ubuntu run:

git clone https://github.com/McDonnell-Lab/DCASE2020_Task1b.git

Open a terminal and run:

bash requirements.sh s This will do the following:

• create a new anaconda environment
• download required python libraries
• download the development dataset (not the evaluation dataset)

The script makes use of this tool: https://gitlab.com/dvolgyes/zenodo_get for downloading.

After this has finished, run

conda activate DCASE2020

To inspect and download data manually:

-download Task 1b data from here: https://zenodo.org/record/3670185#.XmW-ehdLfUp

Training and validation

• first run Task1b_make_spectrograms.ipynb
• next try running jupyter notebook Source/Notebooks/Task1b_for_Github_train_and_validate.ipynb
• for full "leave one city out" (LOCO) cross-validation, run jupyter notebook Source/Notebooks/Task1b_final_design_LOCO.ipynb

Verification that trained 500KB model works

• run jupyter notebook Source/Notebooks/Task1b_saved_model_1_bit_per_weight_storage.ipynb
• The notebook shows a validation accuracy of 96.87% and 486.6943359375 KB for convolutional weight storage
• The notebook creates a .mat file that contains all the binary weights that uses 592 KB on disk (there is about 100KB of overheads stored in it).